3750 performance problem

Unanswered Question
Jun 30th, 2009

We are having a bizarre problem with the performance of a diagnostic imaging application when the client PC is connected to a Catalyst WS-C3750-48TS which is uplinked by Gig fibre to our core Catalyst 6513. What's bizarre is that the application perfroms well when the PC is connected to a Catalyst 2950 uplinked by GigE or by FastE to the same core (see attached diagram).

With Wireshark, we captured traffic from the server's perspective on a span'd switch port on the 6513 and ran two tests... one with the client PC connected to a 2950 and then one witht he PC connected to the 3750. During each test, the application was run in the same manner, generating the same data transmission to the client. In the case of the 3750, part way through application run, we began to see "TCP Window Full" messages followed by a flury of "TCP Dup Ack" , "TCP Fast Retransmission" messages in the Wireshark capture. In contrast however, at the same point in the application run with the PC connected to the 2950, we only observed the "TCP Window Full" messages.

Furthermore, we upgraded the 3750 to the latest code (12.2.50) avaialble, and wiped the switch using it as a basic access switch with the uplinkport configured as an access port.

Something is obviously wrong with the way the 3750s are handling this flow of traffic as opposed to 2950s.

Any insight you can provide would be most helpful as we have been scratching heads on this for some time now.

I have this problem too.
0 votes
  • 1
  • 2
  • 3
  • 4
  • 5
Overall Rating: 5 (2 ratings)
Joseph W. Doherty Tue, 06/30/2009 - 16:14

From what you describe, my guess would be egress buffer on the 3750 might be dropping packets prematurely. I recall somewhere(?) I've read that 3750s were setting their default egress buffers too small although I also thought that had been corrected in later code.

I did find this in the latest release notes "If a switch is forwarding traffic from a Gigabit ingress interface to a 100 Mb/s egress interface, the ingress interface might drop more packets due to oversubscription if the egress interface is on a Fast Ethernet switch (such as a Catalyst 3750-24TS or 3750-48TS switch) than if it is on a Gigabit Ethernet switch (such as a Catalyst 3750G-24T or 3750G-24TS switch). There is no workaround. (CSCed00328) ", but I'm not sure if that's what I read before.


Some things to try: if QoS enabled on 3750, disable; if QoS not enabled, enable

If QoS enabled, adjust queue-set parameters, to try to insure there's sufficient buffer space to handle RWIN.

John Rumball Fri, 07/03/2009 - 10:12

Thanks for your reply, Joseph. This is extremely helpful.

We will likely be contacting TAC regarding this to see what our options are.



This Discussion